US official says Kimi K3 copied Fable; AI researchers are skeptical

Kratsios wrote that “large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable,” amid reported discussions about banning Chinese open-weight models. He did not share details on the sources of his allegations, and Moonshot did not respond to questions about its training process. Treasury Secretary Scott Bessent echoed the claim, saying regulators are finding watermarks of US large language models on many Chinese models — though it isn’t clear what those watermarks consist of, and the Treasury Department did not respond to a query.

AI researchers are doubtful that distillation — the process of querying a model to copy its capabilities — explains Kimi K3’s strength. Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, said a model this strong couldn’t come from distillation alone so soon after Fable’s release, since Fable has only been publicly available since July 1st and there hasn’t been time to distill enough data, train a model, and ship it in two weeks. Nathan Lambert, an AI researcher at the Allen Institute for AI, said distillation is becoming less impactful as Chinese models near the frontier and shift toward reinforcement learning, a costlier technique: large RL runs can require tens of millions of agents, and running that volume through a frontier lab’s API would be prohibitively expensive and possibly too slow to explain Kimi K3’s quick turnaround.

The allegation follows Anthropic’s own claim earlier this year that Moonshot, DeepSeek, and MiniMax systematically distilled its models, based on millions of exchanges it traced through IP addresses and other metadata, which it said pointed to deliberate attempts to extract its models’ capabilities rather than ordinary use. Distillation is common industry-wide — Elon Musk has testified that his company SpaceXAI distilled OpenAI models to build Grok, and called the practice common — and the line between distillation and building synthetic datasets can blur.

Kratsios also alleged Moonshot obtained export-restricted Nvidia Grace Blackwell 300 chips and accessed GB300-equipped servers in Thailand. Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology, said a black market for those chips exists and called for “know your customer” reporting requirements at data centers, noting exporters are already supposed to ensure the chips they ship abroad go only to approved uses. In May, the founder of Supermicro, a US server builder, was indicted for smuggling advanced chips into China. The Biden administration’s Commerce Department proposed federal know-your-customer rules for data centers in 2024, but no further progress appears to have been made since.

Why it matters

The dispute shows how hard it is to prove how a frontier model was actually built, and how IP-theft allegations are becoming entangled with chip-export enforcement in the broader US-China AI race.